GNGTS 2023 - Atti del 41° Convegno Nazionale

Session 1.1 - POSTER GNGTS 2023 and test sets, respectively. We used Dataset_2 as a second test set. We first centered the 1.6-seconds-long time windows of training, validation and test sets on the P-wave arrival. The performance of the network on the two test sets is shown in Tab. 2. We reach an overall accuracy on the test set derived from Dataset_1 of 98.9% and an accuracy on Dataset_2 of 95.7%. To simulate the performance in case of uncertainty in P-wave arrival times, we tested the model behavior shifting each time window of test set by a constant value of T samples, with different values of T in the range . The performance is shown in Fig. 3, represented by red lines. It [− 30, 30] is evident that the accuracy rapidly decreases, by at least 20% when the traces are shifted of 10 samples (0.1 seconds). If the translation exceeds 20 samples, the CNN is no longer useful (accuracy less than 50%). Further tests reveal that varying the size of kernels is not helpful in improving performance in the presence of uncertain P-wave picking. Fig. 3 – The performances on the test set of the CFM network after the two different training strategies, i.e. with (in blue) and without (in red) random time shifts in the training set. Performance is shown as a function of the different shift T in the test set. Dashed black lines refer to accuracy levels of 0.5 and 0.75. We performed a second training, including a time-shift in the training set. We perturbed the center of each time window contained in the training set with a uniform random time-shift between -10 and 10 samples. We compared the performances of both training strategies, testing on the same uniformly shifted traces, as shown in Fig. 3. Blue lines refer to the training strategy using time-shift in the training set. In this case performance degradation occurs when . In the case | | > 10 , the performance is approximately constant. Unlucky, when , introducing | | < 10 = 0 time-shift in the training set lowers the performance. A potential reason is that the arrival times in the test set were accurate, and the time-shift was not required.